Abstract
Tensor dimensionality reduction is one of the fundamental tools for moderndata science. To address the high computational overhead, fiber-wise sampledsubtensors that preserve the original tensor rank are often used in designingefficient and scalable tensor dimensionality reduction. However, the theory ofproperty inheritance for subtensors is still underdevelopment, that is, how theessential properties of the original tensor will be passed to its subtensors.This paper theoretically studies the property inheritance of the two key tensorproperties, namely incoherence and condition number, under the tensor trainsetting. We also show how tensor train rank is preserved through fiber-wisesampling. The key parameters introduced in theorems are numerically evaluatedunder various settings. The results show that the properties of interest can bewell preserved to the subtensors formed via fiber-wise sampling. Overall, thispaper provides several handy analytic tools for developing efficient tensoranalysis